ForensicFormer: Hierarchical Multi-Scale Reasoning for Cross-Domain Image Forgery Detection

TL;DR

Cross-domain forgery detection remains challenging as AI-generated imagery evades traditional forensic cues. ForensicFormer integrates three parallel feature streams (low-level frequency/noise, mid-level boundary, high-level semantic plausibility) with hierarchical cross-attention fusion and a triad of task heads (classification, localization, manipulation-type) trained through staged and adversarial strategies. It achieves a cross-domain average accuracy of $86.8\%$ across seven datasets, $0.76$ F1 for pixel-level localization, and robust performance under aggressive JPEG compression ($83\%$ at $Q=70$) while providing interpretable attention maps that align with forensic reasoning. By bridging classical image forensics and modern deep learning, the approach offers a scalable, explainable solution for real-world deployment where manipulation techniques evolve.

Abstract

The proliferation of AI-generated imagery and sophisticated editing tools has rendered traditional forensic methods ineffective for cross-domain forgery detection. We present ForensicFormer, a hierarchical multi-scale framework that unifies low-level artifact detection, mid-level boundary analysis, and high-level semantic reasoning via cross-attention transformers. Unlike prior single-paradigm approaches, which achieve <75% accuracy on out-of-distribution datasets, our method maintains 86.8% average accuracy across seven diverse test sets, spanning traditional manipulations, GAN-generated images, and diffusion model outputs - a significant improvement over state-of-the-art universal detectors. We demonstrate superior robustness to JPEG compression (83% accuracy at Q=70 vs. 66% for baselines) and provide pixel-level forgery localization with a 0.76 F1-score. Extensive ablation studies validate that each hierarchical component contributes 4-10% accuracy improvement, and qualitative analysis reveals interpretable forensic features aligned with human expert reasoning. Our work bridges classical image forensics and modern deep learning, offering a practical solution for real-world deployment where manipulation techniques are unknown a priori.

Figures (4)

• Figure 1: ForensicFormer Architecture. Three parallel branches extract features at different abstraction levels: (a) Low-level branch applies DCT, DWT, and SRM filters to capture frequency and noise artifacts; (b) Mid-level branch uses edge detection and semantic segmentation to identify boundary inconsistencies; (c) High-level branch performs shadow/reflection/depth analysis for physical plausibility. Features are encoded via transformer blocks, then fused via cross-attention. Three output heads predict classification, localization mask, and manipulation type.
• Figure 2: Attention Visualizations (Schematic). (a) GAN-generated face: Low-level branch highlights spectral anomalies (red=high attention). (b) Spliced landscape: Mid-level branch detects boundary discontinuities (orange=high attention). (c) Diffusion-generated scene: High-level branch flags physically implausible shadows (green=high attention). Our hierarchical reasoning aligns with human forensic expertise.
• Figure 3: Failure Cases (Schematic). (a) False Negative: Midjourney v6 portrait with perfect semantic consistency and no technical artifacts (model confidence: 52% real). (b) False Positive: Authentic photo with unusual lens flare triggers low-level detector (model confidence: 72% fake). Future work will address edge cases via uncertainty quantification.
• Figure 4: Accuracy vs. Inference Time Tradeoff. ForensicFormer achieves highest cross-dataset accuracy (86.8%) with reasonable inference cost (500ms). The 3$\times$ slowdown vs. Xception is justified by +11.3% accuracy gain, crucial for forensic applications where accuracy outweighs speed.